What I've been teaching

A seminar for SaaS technical sales and solution engineers, framed around two questions. First: what is vibe coding, really? Second: if AI makes implementation dramatically cheaper, where does the SaaS industry go next? The goal is not to celebrate AI as magic or to declare developers obsolete. The goal is to give field teams a clear agenda for talking about AI, software, and SaaS value when customers increasingly ask, "Why buy this if we can ask AI to build it?"

My view is that vibe coding is best understood as the next abstraction layer. Assembly translated machine operations, programming languages translated those operations into human-ish syntax, and LLM tools now translate natural language into code. That collapses the cost of coding, but it does not collapse the value of software engineering. The scarce work moves from typing implementation to deciding what should exist, why it should exist, which tradeoffs are acceptable, and how the system will survive maintenance, security, regression, cost pressure, and organizational ambiguity.

That is also the agenda for SaaS. If implementation is no longer the primary barrier, SaaS cannot defend itself by saying "we can build screens and workflows." Its value has to be in the operating system around the implementation: domain definitions, standard workflows, trust, uptime, compliance, ecosystem leverage, upgrade paths, partner extensibility, and the ability to turn many customers' repeated problems into a shared product. In other words, the industry has to shift the conversation from build-versus-buy as code production to build-versus-buy as risk, time, quality, and accumulated domain knowledge.

The hands-on sections exist to make that thesis concrete. Claude Code can build a clean dashboard, but it will count open pipeline as revenue unless someone defines revenue as Closed Won. It can build a Tableau-style UI over an API, but it will quietly aggregate 100 rows out of 15,000 unless someone understands the data contract. The final maze asks participants to design an Opus-powered agent that explores APIs, keeps state, and resists traps. The point is the same throughout: AI can implement quickly, but agenda-setting, definition, architecture, and judgment are where technical sales and SEs have to become sharper, not less relevant.

Most "AI agent" lectures start at LangChain. This one starts at a single LLM call and earns every abstraction it adds. Ten chapters, each driven by a real limit that forces the next pattern: from stateless calls → multi-turn → tools → fixed workflows → autonomous tool selection → feedback loops → planning → multi-agent orchestration. The structure is the lesson.

The whole deck is hand-built — reveal.js with a custom warm-ink design system, SVG diagrams with draw-in animations, chapter-tinted accents. No theme bought, no template borrowed. Open-sourced under MIT so others can fork it and adapt it for their own training.

A two-day intensive built for engineers who write firmware, run fabs, design chips — people who'd never touched an LLM and didn't need to become Python developers, but who'd be using AI agents in their workflow within a year. The curriculum bridges that gap. Less code than the GenAI Agent course; more on building intuition about what these systems can and cannot do.

Training infrastructure I also built end-to-end: a React slide SPA, a FastAPI server doing OIDC-SSO into the corporate identity, and an automated LLM-grading pipeline for the exercises. Full curriculum open-sourced under MIT.

A one-shot seminar for an audience that didn't write code. The question on the table: AI is supposedly replacing developers — is it? And if not, what is it actually doing? I made the case for vibe coding as the most visible productivity pattern AI has produced so far, and what it does and doesn't change about who is needed.

Slides are HTML, designed as an editorial keynote rather than a tech deck. Built in the same hand-rolled style as the GenAI Agent course — no framework theme.